Penaeus orientolis prawn freshness rapid determination method based on electronic nose and non-linear stochastic resonance technique

Abstract

In this paper, Penaeus orientolis prawn freshness rapid determination method using electronic nose (e-nose) and non-linear data processing technique is studied. E-nose responses to prawns stored at 4°C are measured. Meanwhile, physical/chemical indexes (firmness, pH, total volatile basic nitrogen (TVB-N), total viable count (TVC), and human sensory evaluation) are examined to provide freshness references for e-nose analysis. E-nose measurement data is analyzed by principal component analysis (PCA), stochastic resonance (SR), and double-layered cascaded serial stochastic resonance (DCSSR). PCA partially discriminates prawns under different storage time. SR and DCSSR signal-to-noise ratio (SNR) spectrum eigen values discriminate prawns successfully. Multi-variables regressions (MVR) are conducted between physical/chemical indexes and SR/DCSSR output SNR minimal (SNR-Min) values. Results indicate that SNR-Min values present more significant linearity relation with physical/chemical indexes. Prawn freshness forecasting model is developed via Harris fitting regression on DCSSR SNR-Min values. Validating experiments demonstrate that forecasting accuracy of this model is 94.29%.

Keywords:

• electronic nose
• freshness
• non-linear stochastic resonance
• penaeus orientolis prawn
• rapid determination

Introduction

Penaeus orientolis prawn is particularly rich in protein, essential amino acids, and essential minerals (Fe, Zn, Se, etc.). It is now the most important shrimp cultivated in the world in terms of volume. It contains rich nutrients such as protein and micronutrients necessary for human body. However, it is hard to preserve and usually deteriorates in a few days. And protein, the main component of white shrimp, will produce toxic amine compounds such as trimethylamine, etc., after deterioration. The intake of deteriorated prawns brings great harm to human body, which has been one of the most serious problems in food safety area. Traditional aquatic products quality analysis methods mainly include instrumental analysis, physical/chemical examination, and human sensory evaluation. Instrumental analysis methods, such as VIS/NIR spectroscopy, gas chromatography-mass spectrometry (GC-MS), and impedance spectroscopy, are usually utilized for aquatic products freshness detection under laboratory conditions.^1-3 However, these methods have some disadvantages, such as high cost, time-consuming, etc. Moreover, skilled operators are required to perform instrumental analysis. Physical/chemical examination methods, such as conductivity/pH measurement and TVB-N, can examine freshness of aquatic products according to relative standards.^4,5 However, these methods present some disadvantages, such as high cost, time-consuming, low reproducibility, etc. Moreover, skilled operators are required to perform the instrumental analytical experiments. So a convenient, fast and robust method is badly needed for prawn freshness determination.

E-nose, imitating human olfactory sense, is a kind of electronic instrument.^6,7 In the past 2 decades, e-nose has been successfully applied in diverse fields such as food quality analysis, environmental control, industrial applications, and clinical diagnostics. Pattern recognition (PR) methods used by e-nose can be divided into supervised and non-supervised methods although a combination of both can be used.^8,9 The major supervised techniques are linear discriminant analysis (LDA), soft independent modeling of class analogy (SIMCA), partial least squares discriminant analysis (PLS-DA), k-nearest neighbors (KNN), artificial neural networks (ANN), and classification and influence matrix analysis (CAIMAN). Unsupervised techniques include cluster analysis (CA), artificial neural networks (ANN) and PCA.¹⁰

In 2001, a portable electronic nose was used to detect fish freshness by estimating the number of storage days.¹¹ An electronic nose system was built based on metal oxide semiconductor (M.O.S) sensors to evaluate freshness of sardines under cold storage.¹² Results showed that the sensor signals were well correlated with the number of days that sardines had undergone. Discriminant factor analysis (DFA) and fuzzy ARTMAP neural networks were used to build classification models. Limbo et al. applied chemical methods and electronic nose to predict the decay and shelf life of European sea bass.¹³ A new optoelectronic nose was developed to monitor shelf-life of fresh sea bream under cold storage.¹⁴ In 2004, this group evaluated freshness of Atlantic salmon using optoelectronic nose.¹⁵ Hui et al. predicted grass carp quality by an e-nose combined with non-linear SR.⁷ The system presented good predicting accuracy for grass carp samples.

In this paper, prawn freshness rapid analysis method based on e-nose technique and non-linear SR is studied. Prawns stored at 4°C under different storage time were measured by e-nose. Physical/chemical indexes (firmness, pH, TVB-N, TVC, and human sensory evaluation) are examined in order to characterize prawns’ freshness. E-nose measurement data is processed by PCA, SR, and DCSSR. MVR results demonstrate that DCSSR eigen values present more significant linearity relation with physical/chemical indexes. Prawn freshness forecasting model is developed based on Harris fitting regression of DCSSR SNR-Min values. Validating experiments demonstrate that the developed model forecasts prawn freshness with an accuracy of 94.29%.

Results and Discussions

Firmness

Prawn firmness examination results are displayed in Figure 1a. Firmness increases slightly in day 2. After day 2, firmness keeps declining and reaches less than 70N in day 6 finally. A possible explanation for this phenomenon lies in the changes of ATP content within shrimps during cold storage. ATP determines the state of actin and myosin. After prawns’ death, glide between actin and myosin happens because of ATP decomposition with the increase of storage days, which leads to muscle stiffness. After reaching stiffness maximal value, endogenous proteases within muscles and exogenous proteases induced by spoilage bacteria make Z line of myofibrillar more weak or broken. So prawns’ muscle becomes soft, which leads to the decline in firmness.

Figure 1. Physical/chemical index examination results: (a) firmness; (b) pH; (c) TVB-N; (d) TVC; (e) sensory evaluation.

pH

Prawn's inner chemical environment changes with the decrease of its quality during storage. pH measurement results are displayed in Figure 1b. pH value is about 7.2 in day 0, and then declines in day 1. From day 2 to day 8, pH keeps increasing. During storage, prawns undergo anaerobic glycolysis procedure. Glycogen degradation and lactic acid formation lead to the rapid decline of pH. When the prawns get deteriorated, trimethylamine oxide within the samples is reduced to trimethylamine with the help of microbes. The formation of trimethylamine leads to slight decline of pH. During further storage, protein within prawns decomposes and releases alkaline substances (ammonia and amines, etc.), which increases pH from day 2 to day 8.

TVB-N

The essential of prawn freshness changes is that the protein decomposes into various chemical substances and metabolites by function of microorganisms. TVB-N index effectively characterizes this decomposing procedure. TVB-N examination results are shown in Figure 1c. TVB-N increases gradually with the increase of storage days. The contents of TVB-N in the first 3 days are 16.50 mg/100 g, 18.60 mg/100 g, and 19.20 mg/100 g, respectively, which accords to China national standard (<20.00 mg/100 g).¹⁶ The results demonstrate that prawns from day 0 to day 2 are fresh and edible. However, TVB-N content of prawns in day 3 is about 20.10 mg/100 g, which exceeds GB/T 5009.44-2003. After day 3, the content of TVB-N increases dramatically, and reaches 39.00 mg/100 g in day 8. The results indicate that prawns from day 3 to day 8 are not fresh.

TVC

Microbial propagation is a crucial factor influencing quality of aquatic products. During storage, prawns deteriorate gradually due to microbiology propagation. Microbiological metabolism leads to emission of feature gases, which provides measurement basis of e-nose analysis. So it is important to trace the microbiological propagation on prawns. TVC values of samples are monitored and displayed in Figure 1d. Initial TVC value is about 2.1×10⁴ cfu•g⁻¹. The TVC value increases dramatically fast after day 3. TVC value in day8 is about 2.2×10⁷ cfu•g⁻¹.

Human sensory evaluation

Sensory attributes of prawn samples were divided into 5 elements, whose preference scores were scored from 1 to 5; the higher preference score, the better quality of prawn (see Table 1). All samples started with score of 5 in day 0.

Table 1. Sensory scheme for evaluating the jumbo squid during the refrigerated storage

	Attribute degree
Attributes	5	4	3	2	1
Appearance of the skin	Very bright	Bright	Slight dull	Dull	Very dull
Touch	Very hard	Hard	Slight soft	Soft	Very soft
Color	Pink	Slight red	Red	Dull red	Black red
Odor	Neutral	Slight fishy	Fishy	Slight odorous	Pungent odorous
Viscidity	Smooth	General smooth	Slight viscous	Viscous	Very viscous

In Figure 1e, preference scores decrease with the increase of storage days. Choosing 3.5 as the acceptable limitation, the prawns can maintain its shelf life by 3 days, and the preference score in day 3 is about 3.39. Sensory evaluation results accord to TVB-N results well. Combined with TVB-N and TVC examination results, sensory evaluation results might be affected by microbial activities and changes of prawns’ self components.

E-nose analysis results

Original responses and PCA results

In e-nose measurement, gas sensors are subjected to time drift. E-nose's stability is evaluated by conducting blank sample measurement. Zero gas, obtained by filtering air through active carbon, is used as blank sample. E-nose responses to blank sample are measured for 7 days. Blank sample measurement results demonstrate that e-nose responses to zero gas remain stable. Meanwhile, the sensors’ time drift during measurement has no effect to e-nose measurement. E-nose responses prawns are recorded. Figure 2a shows e-nose gas sensor responses on day 7. With the increase of measurement time, 8 sensors’ responses increase gradually. The responses of the sensors to sample's headspace gases are similar. Obvious increase occurs in the time interval between 2.5 s and 25 s. Finally, the responses trend to be stable. Sensor S4, S1, S5, S7, S6, S8, and S3 show high responsiveness to all the samples. S4 shows the largest responses. Responses of S2 are much smaller. Gas sensors within e-nose present various responses to prawns, which constructs the analysis basis of e-nose.

Figure 2. E-nose detecting results: (a) prawn sample measurement; (b) PCA result: PC1 VS. PC2; (c) PCA result: PC1, PC2 VS. PC3; (d) SR SNR spectrum; (e) DCSSR SNR spectrum; (f) freshness predicting model.

Figure 2b presents the PCA results. The first 2 principal components, PC1 and PC2, capture 94.5% of data variance. The figure shows the projections of PCA results on a 2-dimensional plane formed by the first 2 principal components. Combined with TVB-N examination results, prawns can be classed into 2 groups: fresh group and non-fresh group. Fresh group corresponds to the samples having undergone up to day 2. Non-fresh group corresponds to the samples from day 3 to day 6. From day 0 to day 2, the samples gather in the first and forth quadrant. Day 3 is the turning point. From day 3 to day 8, PC1 decreases quickly, and PC2 first decrease from day 3 to day 4. From day 4 to day 5, PC2 dramatically increases. In day 6, PC2 decrease slightly. However, Fresh and non-fresh samples have no obviously discriminating border. The samples under different storage time within the same group cannot be clearly discriminated from each other. PCA only provides a qualitative analysis on e-nose measurement data. PCA results on a 3-dimensional plane formed by the first 3 principal components are displayed in Figure 2c. The first 3 principal components, PC1 and PC2, capture 97.44% of data variance. All prawns under different storage time still cannot be totally discriminated from the 3-dimensional angle.

SR and DCSSR SNR spectrum

Prawn e-nose measurement data SNR spectrum calculated by SR as function of external white noise intensity is displayed in Figure 2d. A derivative vale arises at noise intensity of 37. Then SNR increases gradually with the increase of external noise intensity. Eigen peak appears at noise intensity of 65. The SNR maximal (SNR-Max) values range between −82 and −79 dB. SNR-Max values increase with the increase of storage time. Prawn samples under different storage days can be discriminated by SNR-Max values.

Prawn e-nose measurement data SNR spectrum calculated by DCSSR as function of noise intensity are displayed in Figure 2e. A derivative vale arises at noise intensity of 37. SNR increases gradually with the increase of stimulating noise intensity. A derivative peak appears at noise intensity of 74. Then SNR gradually decreases, and reaches its minimal value at noise intensity of 118. SNR-Min values range between −74 and −53 dB. SNR-Min values decrease with the increase of storage time. Prawns of different storage time can be clearly discriminated according to SNR-Min values. Compared with SR output SNR spectrum, DCSSR method especially improves output signal intensity, which contributes to prawn discrimination using SNR eigen values.

MVR between physical/chemical indexes and SR/DCSSR eigen values

Quality is a general description for prawns. The physical/chemical indexes, such as TVB-N, TVC, pH, firmness, have close relationship to prawns’ quality. MVR results between physical/chemical indexes and SR SNR-Max values are displayed as Equation (1)(1) $\begin{array}{l}{\mathrm{(R}}^2=\text{0}\text{.99741)}\hfill \\ \begin{array}{l}SR=-79.5226+0.00102\hfill \\ \times Firmness-0.03659\times pH\hfill \end{array}\hfill \\ \begin{array}{l}+0.08123\times TVBN-2.96\times {10}^{-8}\hfill \\ \times TVC-0.34408\times Sensory\hfill \end{array}\hfill \end{array}$(1) and Table 2. ${\displaystyle R^2=\text{0.99741}}$ indicates that SR eigen values have good linearity with prawn quality indexes. P < 0.05 and F = 77.1599, which indicate that SNR-Max values present significant linearity relation with physical/chemical indexes.(1) $\begin{array}{l}{\mathrm{(R}}^2=\text{0}\text{.99741)}\hfill \\ \begin{array}{l}SR=-79.5226+0.00102\hfill \\ \times Firmness-0.03659\times pH\hfill \end{array}\hfill \\ \begin{array}{l}+0.08123\times TVBN-2.96\times {10}^{-8}\hfill \\ \times TVC-0.34408\times Sensory\hfill \end{array}\hfill \end{array}$(1)

Table 2. Multiple variable regression between SR eigen values and physical/chemical indexes (F= 77.1599)

Index	Value	Error	t-Value	P
Y-Intercept	−79.5226	7.10593	−11.191	0.04674
Firmness	0.00102	0.00732	0.13966	0.91166
pH	−0.03659	0.82895	−0.04414	0.97192
TVB-N	0.08123	0.04032	2.01448	0.29333
TVC	−2.96E-08	1.81E-08	−1.63267	0.34986
Sensory	−0.34408	0.17609	−1.95395	0.30114

MVR results between physical/chemical indexes and DCSSR SNR-Min values are displayed as Equation (2)(2) $\begin{array}{l}\begin{array}{l}DCSSR=-48.29699-0.03516\hfill \\ \times Firmness-0.00878\times pH\hfill \end{array}\hfill \\ \begin{array}{l}+0.02929\times TVBN-2.22\times {10}^{-7}\hfill \\ \times TVC+1.58907\times Sensory\hfill \end{array}\hfill \\ (R^2=\text{0}\text{.99818})\hfill \end{array}$(2) and Table 3. ${\displaystyle R^2=\text{0.99818}}$ demonstrates that DCSSR eigen values have better linearity with prawn quality indexes. P < 0.001 and F = 109.66052, which show that DCSSR SNR-Min values have more significant linearity relation with physical/chemical indexes than SR SNR-Max values. DCSSR SNR-Max values are more suitable for prawn freshness characterization.(2) $\begin{array}{l}\begin{array}{l}DCSSR=-48.29699-0.03516\hfill \\ \times Firmness-0.00878\times pH\hfill \end{array}\hfill \\ \begin{array}{l}+0.02929\times TVBN-2.22\times {10}^{-7}\hfill \\ \times TVC+1.58907\times Sensory\hfill \end{array}\hfill \\ (R^2=\text{0}\text{.99818})\hfill \end{array}$(2)

Table 3. Multiple variable regression between DCSSR eigen values and physical/chemical indexes (F= 109.66052)

Index	Value	Error	t-Value	P
Y-Intercept	−48.29699	39.80333	−1.21339	0.00081
Firmness	−0.03516	0.041	−0.85753	0.54873
pH	−0.00878	4.64332	−0.00189	0.9988
TVB-N	0.02929	0.22586	0.12969	0.91789
TVC	−2.22E-07	1.01E-07	−2.18471	0.27328
Sensory	1.58907	0.98637	1.61102	0.35365

Freshness predicting model

Figure 2f shows the individual SNR-Min value of samples under different storage time. Prawn freshness forecasting model is established by non-linear Harris fitting regression of SNR-Min values. The fitting results are displayed as Equation (3)(3) ${\displaystyle y=\frac{1}{0.00004x^{2.4578}-0.02262}, R^2=0.99375}$(3) , and the regression coefficients R² 0.99375. After one-step's transform, Equation (4)(4) ${\displaystyle Freshness=\sqrt[2.4578]{\frac{1+0.02262\times SNRMin}{0.00004\times SNRMin}}}$(4) is used as freshness forecasting model. The input is e-nose measurement data DCSSR processed SNR-Min values, and the output is prawns’ prediction of storage time. According to Equation (4)(4) ${\displaystyle Freshness=\sqrt[2.4578]{\frac{1+0.02262\times SNRMin}{0.00004\times SNRMin}}}$(4) and physical/chemical examination results, we can achieve approximate quality of the measured samples. This method is suitable for most of samples following prawn freshness changing discipline. It is not suitable for quality forecasting of special samples.(3) ${\displaystyle y=\frac{1}{0.00004x^{2.4578}-0.02262}, R^2=0.99375}$(3) (4) ${\displaystyle Freshness=\sqrt[2.4578]{\frac{1+0.02262\times SNRMin}{0.00004\times SNRMin}}}$(4)

Validation experiment results

Validating experiments are carried out to examine predicting accuracy of the developed model. Equation (4)(4) ${\displaystyle Freshness=\sqrt[2.4578]{\frac{1+0.02262\times SNRMin}{0.00004\times SNRMin}}}$(4) is used to examine the freshness of prawns. Another 84 new prawns are prepared, and 35 samples are randomly selected for e-nose measurement. Results are displayed in Table 4. The forecasting accuracy of the predicting model is 94.29%. Validating experiment results demonstrate that the prawn forecasting model detects the prawns’ freshness with good accuracy.

Table 4. Freshness forecasting results (√ = right; x = wrong; / = not calculated)

Sample No.	Forecasting value	Forecasting time(d)	Practical time(d)	Predicting error (%)	Freshness	Result
20	−46.3619	3.78	4	5.82	non-fresh	√
3	−44.2949	1.07	1	7.0	fresh	√
6	−44.5758	1.87	2	6.95	fresh	√
48	−50.6028	5.68	6	5.63	non-fresh	√
80	−46.9865	4.17	4	4.08	non-fresh	√
11	−44.2087	0.04	0	/	fresh	√
17	−45.5388	3.13	3	3.33	non-fresh	√
26	−45.2773	2.87	3	3.33	non-fresh	√
44	−44.2708	0.91	1	9.0	/	x
39	−52.0826	6.11	6	1.80	non-fresh	√
8	−51.4335	5.93	6	1.18	non-fresh	√
15	−51.2268	5.87	6	2.21	non-fresh	√
21	−46.8834	4.11	4	2.68	non-fresh	√
69	−45.5496	3.14	3	4.46	non-fresh	√
2	−44.704	2.11	2	5.21	fresh	√
14	−45.6834	3.26	3	7.98	non-fresh	√
27	−45.5823	3.17	3	5.36	non-fresh	√
36	−44.2759	0.94	1	6.0	fresh	√
54	−45.4752	3.07	3	2.28	non-fresh	√
56	−49.1356	5.17	5	3.29	non-fresh	√
51	−51.1928	5.86	6	2.39	non-fresh	√
62	−44.7337	2.16	2	7.41	fresh	√
9	−45.5068	3.1	3	3.33	non-fresh	√
66	−44.2097	0.17	0	/	/	x
73	−44.2949	1.04	1	4.0	fresh	√
76	−44.6867	2.08	2	4.0	fresh	√
63	−44.616	1.95	2	2.50	fresh	√
79	−46.5896	3.93	4	1.78	non-fresh	√
83	−49.3741	5.26	5	4.94	non-fresh	√
28	−48.3296	4.84	5	3.31	non-fresh	√
19	−52.6192	6.25	6	4.17	non-fresh	√
39	−49.2406	5.21	5	4.03	non-fresh	√
37	−46.4967	3.87	4	3.36	non-fresh	√
24	−46.9175	4.13	4	3.15	non-fresh	√
77	−47.1643	4.27	4	6.32	non-fresh	√

Materials and Methods

Prawn samples

Fresh prawns, purchased from Wen’er aquatic market in Hangzhou, were transported to laboratory in ice box. 280 samples were prepared and packaged with sealing membrane and stored at 277 K for 7 days. Each sample weighs about 25 g. In each day, 40 samples were used for sensory preference, firmness, TVB-N, TVC examination, and e-nose measurement.

In validating experiments, 84 new prawns were selected and stored for 7 days. In each day, 12 samples were used for e-nose measurement. All samples were taken randomly.

Reagents and apparatus

NaOH was purchased from Sinopharm Chemical Reagent Co. Ltd., (Shanghai, China). Boric acid was purchased from Wuxi Jingke Chemical Co. Ltd., (Jiangsu, China). HCl was purchased from Hangzhou Chemical Reagent Co. Ltd., (Hangzhou, China). MgO was purchased from Tianjin Bodi Chemical Co. Ltd., (Tianjin, China). Methyl red and Bromocresol green were purchased from Shanghai SSS Reagent Co., Ltd., (Shanghai, China). Deionized water was used in all experiments. All chemicals were of analytically pure grade or better quality.

E-nose system

Prawn samples were measured by a portable e-nose. The structure is displayed in Figure 3a. It includes 3 main parts: data acquisition, modulating and transmitting unit (U1); sensor array and the chamber unit (U2); power and gas supply unit (U3). Gas sensor array consists of 8 metal oxide semiconductors (M.O.S) of different chemical compositions and thickness to provide selectivity toward different gases. The selectivity toward volatile compound classes of M.O.S sensors is indicated by the supplier: S1 (MQ-2, propane), S2 (MQ-3, ethanol), S3 (MQ-4, methane), S4 (MQ-5, propane, butane), S5 (MQ-6, butane), S6 (MQ-7, carbon monoxide), S7 (MQ-8, hydrogen), S8 (MQ-9, methane, carbon monoxide). The sensor response is expressed as sampling voltage (V). The M.O.S sensors rely on changes in conductivity induced by the adsorption of molecules in the gas phase and on subsequent surface reactions. They consist of ceramic substrate coated by metal oxide semiconducting film, and heated by wire resistor. Due to the high temperature (250–500°C), the volatiles transferred to the surface of the sensors are totally combusted to carbon dioxide and water, leading to a change in the resistance. The high temperature avoids water interference and provides sensors fast response and rapid recovery time. Polytetrafluoroethylene (PTFE) material is utilized to fabricate the chamber. Each sensor room is separated, which helps to avoid the cross-influence of the gas flow.

Figure 3. E-nose system: (a) systematic diagram; (b) SR; (c) DCSSR.

Methods

Firmness

TA.XT2i Texture Analyzer (Stable Micro Systems, UK) is used to conduct firmness measurement. Flat cylindrical probe p/5 (5 mm in diameter) was used. TPA mode was accepted. Speed in pre-measurement and after-measurement was 3 mm/s. Measurement speed was 1 mm/s. Compression degree was 50%. Residence time interval was 5 s. Load probe type was Auto-0.2 g. Data collection rage was 200.

pH

In pH measurement, 10 g prawn was smashed by a muller. Deionized water was added into the smashed sample to fully dissolve chemical substances. The sample and water were poured into conical flask. Muller was washed by deionized water, and washing liquid was also poured into conical flask. The mixture in the conical flask was finally diluted to 100 mL, and left to stand for a while. Then the mixture was filtered and measured by PHS-3C pH meter.

TVB-N

In TVB-N examination, magnesium oxide method was used in which contain (ammonia, mono-methyl amine, diethyl amine and tri-methyl-amine and other volatile amine). TVB-N examination was conducted by micro kieldahl apparatus (Hongji Instruments Company, Shanghai, China). The prawn meat was cut by pulper (Yide Electric Appliance Co., Ltd., Guangdong, China). According to China standard protocols GB/T 5009.44-2003 “Method for analysis of hygienic standard of meat and meat products,”¹⁶ 10 g sample was added to the heating flask containing 100 mL deionized water. After 30 minutes, the mixture was filtered. The filtrate was placed into refrigerator. 10 mL filtrate and 5∼6 drops of indicator were mixed and added into a heating flask. 5 mL of mixture and 5 ml magnesium oxide suspension (10 g/L) were added into distiller. After 5 minutes’ distill, hydrochloric acid solution (0.01 mol/L) was used for titration. The TVB-N was determined as follows:(5) ${\displaystyle X=\frac{(V_1-V_2)\times c\times 14}{m\times 5/100}\times 100}$(5)

Where:

• ${\displaystyle X}$ was the TVB-N index in mgN/100 g.

• ${\displaystyle V_1}$ was the volume of hydrochloric acid used in sample titration in mL.

• ${\displaystyle V_2}$ was the volume of control used in titration in mL.

• ${\displaystyle c}$ was the concentration of hydrochloric acid in mol/L.

• ${\displaystyle m}$ was the weight of sample in grams.

TVC

25 g of prawn meat was aseptically weighted and homogenized with 225 ml sterile 0.1% physiological saline for 1 min. The homogenized sample was serially diluted using 9 ml sterile saline for bacteriological analysis. TVC was determined in plate count agar by the spread plate method.¹⁷

Sensory evaluation

According to previously reported method by Amerine et al.,¹⁸ the sensory of prawn samples was evaluated by 6 experienced panelists (2 males and 4 females). Voting number is set at k, k ∈ (1,10). The quality of prawn is divided into m levels, and the score of a specific level is set at h${\displaystyle }$, j ∈ (1, m). Prawn attributes are divided into n elements, and a specific element is set at u_i, i ∈ (1,n). The contributory weight is determined by pairwise comparison of contribution weight of attributes is set at x_i (∑x_i = 1). If there is a specific relationship between 2 objects of h${\displaystyle }$ and u_i, the relation set (matrix) of f is calculated as follows:(6) ${\displaystyle F=\left|\begin{array}{l}{f}_{11}/k{f}_{11}/k\dots f_{1m}/k\\ f_{21}/k{f}_{22}/k\dots f_{2m}/k\\ \dots \dots \dots \dots \\ f_{n1}/k{f}_{n1}/k\dots f_{nm}/k\end{array}\right|}$(6)

Thus, the overall acceptability of prawn is calculated by the weight grade method as follows:(7) ${\displaystyle Z=\sum _{i=1}^{n}xi•\sum _{m=1}^m\frac{fij}{k}hj}$(7)

E-nose measurement

In e-nose measurement, 25 g of samples were placed into 50 mL glass air-tight vials, and sealed with sealing membrane. The vials were stored at 277 K until analysis. The vials were equilibrated at room temperature for 45 min and analyzed in standardized conditions. After the power is turned on, washing pump and valve 2 are started, and sampling pump and valve 1 are close. The air is filtered by active carbon to obtain zero gas. Sensor array are washed by zero gas. When sensors’ response reach the baseline, the washing pump and valve 2 are closed. Then the sampling pump and valve 1 are started. The e-nose sucked the headspace gases of samples through the sensor array at 400 mL/min for 40 s. The measurement interval of e-nose was 0.05 s. When measurement is finished, the sensors are washed by zero gas for 600 s at a flow rate of 1000 mL/min prior to the next sample measurement.

Quality predicting method

SR

SR was proposed by Benzi to give an explanation for Earth climate periodical changes.^19-24 SR model can be described as follows:(8) ${\displaystyle \frac{dx}{dt}=-\frac{dV(x)}{dx}+MI(t)+C\xi (t)}$(8)

Where ${\displaystyle x}$ is the position of the Brownian particle, ${\displaystyle t}$ is the time, M and C are adjustable parameters, ${\displaystyle I(t)=S(t)+N(t)}$ is an input signal ${\displaystyle S(t)}$ with an intrinsic noise ${\displaystyle N(t)}$, ${\displaystyle \xi (t)}$ is the external noise, and ${\displaystyle V(x)}$ is the simplest double-well potential with the constants ${\displaystyle a}$ and ${\displaystyle b}$.(9) ${\displaystyle V(x)=-\frac{1}{2}a{x}^2+\frac{1}{4}b{x}^4}$(9)

Equation (8)(8) ${\displaystyle \frac{dx}{dt}=-\frac{dV(x)}{dx}+MI(t)+C\xi (t)}$(8) can be transformed as(10) ${\displaystyle \frac{dx}{dt}=ax-b{x}^3+MI(t)+C\xi (t)}$(10)

The minima of ${\displaystyle V(x)}$ are located at ${\displaystyle \pm x_m}$, where ${\displaystyle x_m={(a/b)}^{1/2}}$. A potential barrier separates the minima with the height given by ${\displaystyle \Delta U=a^2/4b}$. The barrier top is located at ${\displaystyle x_b=0}$. When three elements of SR interact coherently, the potential barrier can be reduced and the Brownian particle may surmount the energy barrier and enter another potential well. The intensity of signals will increase, which makes it possible that the weak signal can be detected from noise background.

Suppose the input signal is ${\displaystyle I(t)=Asin(2\pi ft+\varphi )}$, where ${\displaystyle A}$ is signal intensity, ${\displaystyle f}$ is signal frequency. ${\displaystyle D}$ is external noise intensity. SNR is the common quantifier for SR and it can be approximately described as:(11) ${\displaystyle SNR=\sqrt{2}\Delta U{(A/D)}^2{e}^{-\Delta U/D}}$(11)

Noise intensity is a parameter in SR model. This model is used for e-nose data analysis. ${\displaystyle I(t)=Asin(2\pi ft+\varphi )+EN(t)+N(t)}$ denotes an input matrix. It has a sinusoid signal ${\displaystyle Asin(2\pi ft+\varphi )}$, electronic nose response data ${\displaystyle EN(t)}$, and intrinsic noise ${\displaystyle N(t)}$. SNR between the output and input is calculated. This model has been successfully used in food analytical applications.^19-22,25 The schematic diagram of SR is showed as Figure 3b.

DCSSR

Under adiabatic elimination condition, supposing signal amplitude is much smaller (A≪1), the Brownian particle is in one of potential wells because no enough driving energy is provided by the bistable system to drive the particle to jump from one potential well to another one in the absence of external noise. The signal period is longer than some characteristic intrawell relaxation time for the system. The existence of periodic forcing inclines the potential function, and forms the Brownian particle's transfer from one potential well to another one. So the potential function ${\displaystyle V(x)}$ changes with the input signal and becomes(12) $\begin{array}{l}V(x,t)=-\frac{1}{2}a{x}^2+\frac{1}{4}b{x}^4+Ax\sin (2\pi ft+\phi )\\ +EN(t)x+N(t)x\end{array}$(12)

Equation (12)(12) $\begin{array}{l}V(x,t)=-\frac{1}{2}a{x}^2+\frac{1}{4}b{x}^4+Ax\sin (2\pi ft+\phi )\\ +EN(t)x+N(t)x\end{array}$(12) indicates the potential function gets time dependence. Equation (13)(13) $\{\begin{array}{l}\begin{array}{l}\partial V(x,t)/\partial x=-ax+b{x}^3+A\sin (2\pi ft+\phi )\\ +EN(t)+N(t)=0\end{array}\hfill \\ \partial V^2(x,t)/\partial x^2=-a+3b{x}^2=0\hfill \end{array}$(13) displays the first-order and second-order derivation of ${\displaystyle V(x,t)}$ with respect to ${\displaystyle x}$, and let the equations equal to zero:(13) $\{\begin{array}{l}\begin{array}{l}\partial V(x,t)/\partial x=-ax+b{x}^3+A\sin (2\pi ft+\phi )\\ +EN(t)+N(t)=0\end{array}\hfill \\ \partial V^2(x,t)/\partial x^2=-a+3b{x}^2=0\hfill \end{array}$(13)

Setting noise intensity ${\displaystyle D=0}$ and ${\displaystyle sin(2\pi ft+\varphi )=1}$, the critical amplitude value of the periodic signal can be obtained: $A_c=\sqrt{4a^3/27b}$. If ${\displaystyle A<A_c}$, the particle hovers around its original stable state and cannot jump from one potential well to another one. However, the particle can jump from one potential well to another one with the help of external noise even if ${\displaystyle A<A_c}$, which means the occurrence of SR phenomenon. We use fourth-order Runge-Kutta numerical algorithm to solve Equation (13)(13) $\{\begin{array}{l}\begin{array}{l}\partial V(x,t)/\partial x=-ax+b{x}^3+A\sin (2\pi ft+\phi )\\ +EN(t)+N(t)=0\end{array}\hfill \\ \partial V^2(x,t)/\partial x^2=-a+3b{x}^2=0\hfill \end{array}$(13) :(14) ${\displaystyle x_{n+1}=x_n+1/6\left[k_1+(2-\sqrt{2})k_2+(2+\sqrt{2})k_3+k_4\right],n=0,1,,N-1}$(14) (15) ${\displaystyle k_1=h\left(a{x}_n-b{x}_n^3+s{n}_n\right)}$(15) (16) ${\displaystyle k_2=h\left[a\left(x_n+\frac{k_1}{2}\right)-b{\left(x_n+\frac{k_1}{2}\right)}^3+s{n}_n\right]}$(16) (17) ${\displaystyle k_3=h\left[a\left(x_n+\frac{k_2}{2}\right)-b{\left(x_n+\frac{\sqrt{2}-1}{2}{k}_1+\frac{2-\sqrt{2}}{2}{k}_2\right)}^3+s{n}_{n+1}\right]}$(17) (18) ${\displaystyle k_4=h\left[a\left(x_n+k_3\right)-b{\left(x_n-\frac{\sqrt{2}}{2}{k}_2+\frac{2+\sqrt{2}}{2}{k}_3\right)}^3+s{n}_{n+1}\right]}$(18)

${\displaystyle x}$ is the nth numerical value of ${\displaystyle x(t)}$, and ${\displaystyle s{n}_n}$ is the nth numerical value of ${\displaystyle S_n(t)}$. ${\displaystyle h}$ is the computation step. Much progress has been achieved on the applications of SR in the past few decades.

Single SR system connected in series forms the cascaded SR to obtain DCSSR (see Figure 3b). According to Equation (10)(10) ${\displaystyle \frac{dx}{dt}=ax-b{x}^3+MI(t)+C\xi (t)}$(10) , the relative Langevin equations of the cascaded bistable systems can be respectively written as:(19) ${\displaystyle \{\begin{array}{l}d{x}_1/dt=a{x}_1-b{x}_1+M[Asin(2\pi ft+\varphi )+EN(t)+N(t)]\\ d{x}_2/dt=a{x}_2-b{x}_2+x_1(t)\\ \dots \\ d{x}_n/dt=a{x}_n-b{x}_n+x_{n-1}(t)\end{array}}$(19)

In practical engineering measurement, measured data usually consists of signal and intrinsic noise. If an aimed weak signal is submerged in strong noise, we are not able to detect it. With the help of SR, the energy of intrinsic noise is lowered, and the embedded weak signal is amplified effectively so that the signal can be caught for measurement characterization. In some special measurement, the weak signal is so weak that the aimed signals are still embedded in the noise. DCSSR is designed to solve this problem. For a DCSSR system, several bistable systems are in serial. If the measurement signals are still embedded in the noise, the output of the former bistable system is taken as the latter bistable system. If the aimed signals are successfully obtained, the current number of bistable systems is the suitable for the information measurement. The schematic diagram of DCSSR is showed as Figure 3c.

Mechanism analysis of DCSSR

From above results, external white noise plays an important part in e-nose measurement data feature extraction. What is acting mechanism between external nose and weak feature signal? How can e-nose signals be effectively strengthened? Here, the weak signal amplification efficiency is theoretically evaluated. According to Equation (9)(9) ${\displaystyle V(x)=-\frac{1}{2}a{x}^2+\frac{1}{4}b{x}^4}$(9) , Equation (8)(8) ${\displaystyle \frac{dx}{dt}=-\frac{dV(x)}{dx}+MI(t)+C\xi (t)}$(8) can be obtained:(20) ${\displaystyle \{\begin{array}{l}enose(t)+N(t)=ax-b{x}^3-Asin(2\pi ft+\varphi )\\ x=\sqrt{a/3b}\end{array}}$(20)

In SR analysis, we suppose the initial intrinsic noise signal ${\displaystyle N(t)}$ is twice the e-nose signal ${\displaystyle enost(t)}$ strength. Then $enose(t_1)\approx \frac{ax-b{x}^3-A_1{sin}_1(2\pi f{t}_1+{\varphi }_1)}{3}$. In DCSSR analysis, if the intrinsic noise signal can be reduced to 0.15 times the e-nose signal ${\displaystyle enose(t_2)}$, then $enose(t_2)\approx \frac{ax-b{x}^3-A_2{sin}_2(2\pi f{t}_2+{\varphi }_2)}{1.15}$. ${\displaystyle \Delta }$ is defined as the signal amplification coefficient, and ${\displaystyle \Delta =enose(t_2)/enose(t_1)}$. After simplification by $x=\sqrt{a/3b}$ and ${\displaystyle A_1=A_2=A_c=\sqrt{4a^3/27b}}$, Equation (21)(21) ${\displaystyle \Delta =\frac{3(1-{sin}_2(2\pi f{t}_2+{\varphi }_2))}{1.15(1-{sin}_1(2\pi f{t}_1+{\varphi }_1))}}$(21) can be obtained:(21) ${\displaystyle \Delta =\frac{3(1-{sin}_2(2\pi f{t}_2+{\varphi }_2))}{1.15(1-{sin}_1(2\pi f{t}_1+{\varphi }_1))}}$(21)

According to Equation (19)(19) ${\displaystyle \{\begin{array}{l}d{x}_1/dt=a{x}_1-b{x}_1+M[Asin(2\pi ft+\varphi )+EN(t)+N(t)]\\ d{x}_2/dt=a{x}_2-b{x}_2+x_1(t)\\ \dots \\ d{x}_n/dt=a{x}_n-b{x}_n+x_{n-1}(t)\end{array}}$(19) , the limits of signal amplification coefficient ${\displaystyle \Delta \approx \infty }$ in theory, which demonstrates that the e-nose signal can be tremendously amplified in practical measurement even if the intrinsic noise ${\displaystyle N(t)}$ is faced to a minor reduction. As a non-linear signal processing method, SR transfers intrinsic noise energy into the weak input signal. So the input weak signal is strengthened and the intrinsic noise is effectively reduced. For prawn samples with different freshness, e-nose feature measurement information is easily submerged into the interfering signals induced by the interactions between other gases and gas sensor array. E-nose feature information gets too weak to be presented in the experiments when using SR. DCSSR analysis relies on the effective combination of 2 or more SR procedures. In this study, 2 SR models are cascaded in serial successively. Intrinsic noise can be tremendously reduced and e-nose feature information can be effectively amplified via DCSSR method.²⁵ Experimental results demonstrate that DCSSR method is more effective than SR method in prawn freshness determination utilizing e-nose.

Conclusions

In this paper, penaeus orientolis prawn freshness rapid determination using e-nose and non-linear SR technique is proposed. The results are listed in the following:

1. E-nose responses to prawn samples stored at 4°C under different storage time are measured. Physical/chemical indexes (firmness, pH, TVB-N, TVC, and human sensory evaluation) are examined to provide freshness references for e-nose analysis.

2. Firmness and sensory preference scores increases with the increase of prawn storage time. However, pH, TVB-N, and TVC indexes increases during experimental procedure. According to standard protocols, that prawns from day0 to day 2 are fresh and the rest samples are not fresh.

3. E-nose measurement data is processed by PCA, non-linear SR and DCSSR methods. PCA cannot discriminate prawns. SR/DCSSR SNR spectrum eigen values successfully discriminate all prawns.

4. MVR regression is conducted between SR/DCSSR output eigen values and physical/chemical indexes. MVR results demonstrate that DCSSR SNR-Min values have more significant linearity relationship with physical/chemical indexes.

5. Penaeus orientolis prawn freshness forecasting model $Freshness=\sqrt[2.4578]{\frac{1+0.02262\times SNRMin}{0.00004\times SNRMin}}$ is developed via non-linear Harris fitting regression on DCSSR SNR-Min values. Validating experimental results demonstrate that the forecasting accuracy of the developed predicting model is 94.29%.

Aquatic products contain rich nutrition, but they deteriorate easily. According to food quality examination standards in operation, the analysis procedure is fussy and time-consuming. We are carrying further investigations on applications of e-nose technique in food quality rapid analysis.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Funding

This work is financially supported by National Spark Technology Project (Grant No. 2013GA700187), National Natural Science Foundation of China (Grant No. 81000645), Higher Education Research Project of Zhejiang Gongshang University (Xgy13080), Student Innovation Projects of Zhejiang Gongshang University (13-157, 13-158, 14-166, 14-193).